Fetching from the wire…
01
02
03
04
05
06
07
08
Infra2026-09-18 · source-backed
llama.cpp PR #28501 found count_experts.comp sizing its shared arrays with BLOCK_SIZE (256), so row-id hoisting was disabled entirely for models with more than 256 experts and every workgroup rescanned the full ids tensor. Qwen3.8-Flash-Next has 512. Raising the limit to a separate MAX_EXPERTS of 512 took per-op time from 14,033µs to 7,508µs on Strix Halo, and Q5_K prefill from 426 to 507 t/s at 8k, with token generation unchanged and greedy output identical. Same day, PR #29036 corrects GGML_QUANT_SIZES[Q8_1] in gguf-py from 40 bytes to the real 36, since block_q8_1 stores its scales as ggml_half, so Python tooling computing offsets from that constant was off by 4 bytes per block.
Each link below shares sources, entities, or timing with this story.
The leaderboard says first place. The methodology says you should check your own bill. Qwen3.8 Max now ranks first on Artificial Analysis' agentic index, scoring 86.1 on OSWorld-Verified ahead of GPT-5.6 Sol Max at 83.2 and Fable 5 at 85.0, priced at $2.00/M input and $6.00/M...
Strix Halo and Strix Point default to Vulkan instead of ROCm for up to 23% faster prompt processing and 8% faster generation, and AMD iGPUs without ROCm move to Vulkan instead of CPU on Linux. On Apple Silicon, gated-delta models train up to 25% faster and quantized MLX KV cac...
Claims up to 2x faster generation, and adds fine-tuning of both MoE models on text or image datasets on Apple Silicon via MLX (release). Follow-up turns in long Qwen chats on Mac are reported up to 30x faster, MLX models now use full context size, and GLM-5.3 MLX fine-tunes ex...
Released August 27 with GGUFs for both, claiming 5x faster inference for RAM offloading, working repeated compaction, chats that recover after disconnects instead of losing the reply, and memory estimates shown before a load (GitHub). That's roughly a 24-hour turnaround from t...
Created August 24, it holds a trendingScore of 3,967 against second-place GLM-5.3-Flash at 1,376 (Hugging Face). The near-1:1 like-to-download ratio means almost everyone bookmarking it hasn't pulled weights, and the unsloth GGUF conversion at 4,354 downloads is absorbing comp...
On a warmed M3 Ultra A/B, Qwen3.8-27B went from 458 tok/s baseline at 4K prefill to 588 with the ANE/GPU split and 625 with ANE/CPU/GPU, costing about 7 GB peak memory. GitHub The compile-memory drop is arguably the bigger deal, since a 35.8 GB spike locked out most Macs from...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.